[E] Add TurboQuant support #450

Epic to track TurboQuant support in SKaiNET.

Goal: make TurboQuant a first-class runtime feature, starting with KV-cache compression.

• new storage/layout abstractions beyond DType
• packed/streamed tensor storage
• KV-cache compression integration at SDPA
• CPU reference path and backend kernels
• loader changes for GGUF/SafeTensors